README.md

Office Workflow OpenEnv Environment

office_workflow_env simulates three realistic office tasks humans do and exposes them via OpenEnv’s standard reset() / step() / state() API.

It supports deterministic grading (0.0–1.0) for:

1. email_triage (easy)
2. data_cleaning (medium)
3. support_escalation (hard)

At the end of each episode, the environment writes the final task score to observation.info.final_score and the structured breakdown to observation.info.final_breakdown.

Task / Episode API

Reset

Call reset(task_id=..., seed=..., episode_id=...).

• task_id: one of email_triage, data_cleaning, support_escalation
• seed: integer seed for deterministic shuffling

Step

Call step(action) where action is an OfficeAction.

The server returns:

• observation: an OfficeObservation
• reward: scalar float (shaped with partial progress)
• done: boolean when the task is finished or max_steps is reached

The info field is carried inside observation.info.

State

Call state() to retrieve OfficeState (progress + action trace; no hidden ground truth answers).

Action Space (OfficeAction)

Single action schema for all tasks, discriminated by type:

• type="triage_email"

  • email_id: string
  • category: one of billing, technical, spam, other
  • priority: int in [0, 5]

• type="correct_cell"

  • row_id: string
  • column: one of email, phone, date
  • value: proposed cleaned value

• type="support_decision"

  • ticket_id: string
  • intent: one of ask_for_information, resolve, escalate
  • reply: draft reply text
  • plus intent-specific fields:
    • requested_info (when ask_for_information)
    • resolution_steps (when resolve)
    • escalation_reason (when escalate)

Optional helper actions:

• type="request_status"
• type="noop"

Observation Space (OfficeObservation)

Common fields:

• task_id, status, max_steps, step_index

Task-specific fields:

• email_triage: emails, triaged
• data_cleaning: dataset, cleaned_cells
• support_escalation: tickets, handled_tickets

Grader-facing info:

• info.objective, info.completion_rule (task instructions)
• info.reward_breakdown (per-step reward component breakdown)
• final episode results in info.final_score and info.final_breakdown

Reward Shaping (partial progress)

Rewards are not binary: they increase as the agent makes correct partial progress.

All tasks:

• progress component grows with corrected/handled items
• correctness component rewards correct submissions
• stalling/wrong actions are penalized by driving reward toward 0.0

Local Setup

Step-by-step (Windows / PowerShell)

1. Open a terminal in c:\Users\hp\Desktop\hackathon
2. Install dependencies:

python -m pip install -U pip
python -m pip install "openenv-core[core]>=0.2.1" openai requests uvicorn

3. Validate OpenEnv structure:

openenv validate

4. Start the server:

uvicorn server.app:app --host 0.0.0.0 --port 8000

5. Confirm it responds:

   • http://localhost:8000/health
   • http://localhost:8000/docs

6. Run the connectivity smoke test (expected score 0.0 because it sends only noop):

python scripts/smoke_test.py

7. Run a “real” local demo baseline (no OpenAI key required; rule-based):

python scripts/baseline_inference.py

Baseline Inference (OpenAI)

The baseline script runs a model against all 3 tasks and prints reproducible scores.

Environment variables:

• OPENAI_API_KEY
• OPENAI_MODEL (optional, default gpt-4o-mini)
• OPENENV_BASE_URL (optional, default http://localhost:8000)
• BASELINE_SEED (optional, default 123)

Run:

python scripts/baseline_inference.py

Hugging Face Spaces Deployment

This repo includes a server/Dockerfile suitable for Hugging Face Spaces.

When you’re ready to deploy:

openenv validate
openenv push --repo-id YOUR_HF_USERNAME/office-workflow-env

openenv push will package the environment and build the Docker image for Spaces.

After it deploys, validate the running Space:

openenv validate https://YOUR_HF_USERNAME-office-workflow-env.hf.space